[PATCH/AARCH64] Fix 32bit reloc for ILP32

[Yufeng Zhang]

Hi Andrew,

Thanks for the patch.

On 07/23/13 23:31, Andrew Pinski wrote:
> Index: bfd/elfnn-aarch64.c
> ===================================================================
> RCS file: /cvs/src/src/bfd/elfnn-aarch64.c,v
> retrieving revision 1.7
> diff -u -p -r1.7 elfnn-aarch64.c
> --- bfd/elfnn-aarch64.c	3 Jul 2013 17:23:24 -0000	1.7
> +++ bfd/elfnn-aarch64.c	23 Jul 2013 22:18:07 -0000
> @@ -344,7 +344,11 @@ static reloc_howto_type elfNN_aarch64_ho
>   	 32,			/* bitsize */
>   	 FALSE,			/* pc_relative */
>   	 0,			/* bitpos */
> +#if ARCH_SIZE == 64
>   	 complain_overflow_unsigned,	/* complain_on_overflow */
> +#else
> +	 complain_overflow_signed,	/* complain_on_overflow */
> +#endif

complain_overflow_unsigned is wrong to use regardless of ELF64 or ELF32, 
but complain_overflow_signed is not the right choice either.  According 
to AAELF64 4.6.5, the 32-bit relocs shall have the overflow check of 
[-2^31, 2^32), and a following note explains that "these overflow ranges 
permit either signed or unsigned narrow values to be created from the 
intermediate result viewed as a 64-bit signed integer. ...".  In other 
words, the 32-bit relocs are only overflowed when the check fails in 
both cases of complain_overflow_unsigned complain_overflow_signed.  FYI, 
there is another checking type - complain_overflow_bitfield, but it is 
not suitable either as it allows some extra negative values.

I suggest to introduce a new check type which checks the range of 
[-2^(N-1), 2^N) for 32-bit and 16-bit relocs (including ABS and PREL). 
For 64-bit relocs, their howto entries should really have 
complain_overflow_dont instead.

The issue doesn't seem to be ILP32-specific.

Regards,
Yufeng